[Paper Review] Hate Speech Detection on Vietnamese Social Media Text using the Bi-GRU-LSTM-CNN Model

Tin Van Huynh, Duc-Vu Nguyen+3 more|arXiv (Cornell University)|Nov 9, 2019

Hate Speech and Cyberbullying Detection15 references32 citations

TL;DR

The paper implements a Bi-GRU-LSTM-CNN classifier for Vietnamese hate speech detection in VLSP 2019, achieving 70.576% F1 on the public test and ranking 5th.

ABSTRACT

In recent years, Hate Speech Detection has become one of the interesting fields in natural language processing or computational linguistics. In this paper, we present the description of our system to solve this problem at the VLSP shared task 2019: Hate Speech Detection on Social Networks with the corpus which contains 20,345 human-labeled comments/posts for training and 5,086 for public-testing. We implement a deep learning method based on the Bi-GRU-LSTM-CNN classifier into this task. Our result in this task is 70.576% of F1-score, ranking the 5th of performance on public-test set.

Motivation & Objective

• Motivate hate speech detection in Vietnamese social media as a low-resource NLP task.
• Develop and compare neural architectures for three-label classification (HATE, OFFENSIVE, CLEAN).
• Demonstrate the effectiveness of a Bi-GRU-LSTM-CNN model on VLSP 2019 data.

Proposed method

• Preprocess text with lowercasing, whitespace/punctuation removal, number replacement, and tokenization using PyVi.
• Represent sentences as 220x300 word embedding matrices using FastText embeddings.
• Apply 1D CNN with dropout followed by Bi-LSTM and Bi-GRU blocks for feature extraction.
• Train a classifier with three labels (HATE, OFFENSIVE, CLEAN) and evaluate using F1-score.
• Compare TextCNN, Bi-GRU-CNN, and Bi-GRU-LSTM-CNN architectures to identify the best model.

Experimental results

Research questions

• RQ1 What is the effectiveness of Bi-GRU-LSTM-CNN for Vietnamese hate speech detection compared to other neural architectures?
• RQ2 How do different neural components (CNN, Bi-LSTM, Bi-GRU) contribute to classification performance on VLSP 2019 data?
• RQ3 What F1-score can be achieved on the VLSP 2019 public test for three-label classification?
• RQ4 How does the model rank on the public-test vs private-test sets?

Key findings

• Bi-GRU-LSTM-CNN yields the best F1-score among the tested models with 70.576%.
• TextCNN and Bi-GRU-CNN achieve 56.512% and 69.293% F1, respectively.
• On the public-test, the Bi-GRU-LSTM-CNN ranked 5th in the top-5 results.
• The approach generally performs comparably to other teams on the public-test but less so on the private-test.
• The dataset used contains 20,345 training and 5,086 test examples with three labels (HATE, OFFENSIVE, CLEAN).